In a significant shift for the high-bandwidth memory (HBM) market, Samsung Electronics is set to leverage TSMC’s cutting-edge 3nm manufacturing process for the base die of NVIDIA’s next-generation custom HBM products. This move marks a departure from Samsung’s previous strategy of producing both the core die and the base die entirely in-house.
Strategic Collaboration for Standardized Performance
Industry sources revealed that NVIDIA has directed its three key NVHBM suppliers—Samsung Electronics, SK Hynix, and Micron—to utilize TSMC’s 3nm process for manufacturing the base die. Concurrently, NVIDIA requires the core die to be produced using its proprietary 1c process, which corresponds to the sixth generation of DRAM technology. These custom HBM solutions are being developed based on the seventh-generation HBM4E architecture and represent NVIDIA’s latest offering in custom high-bandwidth memory, announced in August.
NVIDIA’s directive stems from a desire to standardize performance across different HBM suppliers. Previously, variations in manufacturing approaches by each vendor led to inconsistencies in performance, making it challenging to achieve stable and uniform results. By mandating a unified process for the base die, NVIDIA aims to streamline product validation and ensure a more predictable performance baseline across its supply chain.
Samsung’s Device Solutions division, which oversees its foundry business, has reportedly evaluated and accepted NVIDIA’s request. This collaboration signifies Samsung’s flexibility in adapting to the stringent requirements of major clients like NVIDIA.
Samsung’s Continued In-House Production for Other Products
Despite this new arrangement for NVIDIA’s custom HBM, Samsung plans to maintain its all-in-one, in-house production strategy for its standard HBM offerings. The company recently achieved mass production of the sixth-generation HBM4 and is currently validating the seventh-generation HBM4E for mass production. For these products, Samsung is utilizing its own 4nm foundry process for the base die, with core dies manufactured on its 1c production lines at its facilities in Hwaseong and Pyeongtaek.
Furthermore, Samsung intends to continue its integrated manufacturing approach for custom HBM solutions supplied to other clients, such as AMD. This indicates that the decision to use TSMC for NVIDIA’s base dies is specific to the requirements of that particular partnership and product line.
Contingency Planning for Future Demand
Samsung also remains open to the possibility of developing and producing the NVHBM base dies on its own domestic production lines in the future. Industry insiders suggest that Samsung is actively verifying the feasibility of manufacturing these base dies in-house. This serves as a contingency plan, ensuring that Samsung can scale up production if demand for NVHBM exceeds TSMC’s capacity. This strategic foresight underscores Samsung’s commitment to meeting market demands while exploring optimal manufacturing solutions.
The integration of TSMC’s advanced node for critical components like the base die highlights the intricate interdependencies and strategic alliances shaping the advanced semiconductor manufacturing landscape. As AI and high-performance computing continue to drive demand for HBM, such collaborations are likely to become increasingly common.
